Goal-directed navigation in humans and deep reinforcement learning agents relies on an adaptive mix of vector-based and transition-based strategies

Abstract

Much has been learned about the cognitive and neural mechanisms by which humans and other animals navigate to reach their goals. However, most studies have involved a single, well-learned environment. By contrast, real-world wayfinding often occurs in unfamiliar settings, requiring people to combine memories of landmark locations with on-the-fly information about transitions between adjacent states. Here, we studied the strategies that support human navigation in wholly novel environments. We found that during goal-directed navigation, people use a mix of strategies, adaptively deploying both associations between proximal states (state transitions) and directions between distal landmarks (vectors) at stereotyped points on a journey. Deep neural networks meta-trained with reinforcement learning to find the shortest path to goal exhibited near-identical strategies, and in doing so, developed units specialized for the implementation of vector- and state transition-based strategies. These units exhibited response patterns and representational geometries that resemble those previously found in mammalian navigational systems. Overall, our results suggest that effective navigation in novel environments relies on an adaptive mix of state transition- and vector-based strategies, supported by different modes of representing the environment in the brain.